Discrete Time Series Clustering and Delineation: A Tree-Based Approach to Linear Temporal Logic DiscoveryDownload PDF

13 Apr 2022, 18:45 (modified: 15 Jun 2022, 19:15)XAIP 2022Readers: Everyone
Keywords: Contrastive Explanations, Plan Explanations, Time-series, Planning, Clustering, Decision Trees
TL;DR: In this paper, we introduce a novel discrete time-series clustering and delineation method that embraces the power of linear temporal logic to identify and differentiate unique systems of behaviour.
Abstract: Inferring temporal logic specifications from plan traces can offer significant insight into several aspects of planning such as goal recognition, policy summarization, and system dynamic modelling. Prior work in this area has predominantly focused on the identification of specifications that satisfy all plan traces within a set, however more recently, contrastive approaches concerning the delineation of two sets have also been established. While these approaches are effective in their defined scope, they assume the existence of only one or two behavioural clusters. In this paper, we re-imagine contrastive specification learning by proposing a novel tree generation technique which allows k clusters to be discovered. By embracing a Monte Carlo node-splitting approach, our algorithm seeks balance to contrastively divide any given set of plan traces into two sets with an accompanying temporal logic specification satisfying one of the sets. Recursing this procedure, we demonstrate the effectiveness of our approach to cluster and delineate plan traces, allowing temporal logic specifications to evoke insight at each level of the resulting tree.
7 Replies

Loading